1. Rade Stanojevic Using Tuángòu to Reduce IP Transit Costs Rade Stanojevic (Joint work with Ignacio Castro and Sergey Gorinsky) ‍‍ ACM CoNext 2011 Tokyo, Japan 团购 (tuángòu) = group-buying

2. Rade Stanojevic Part 1: Tuángòu for IP transit ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs2

3. Rade Stanojevic IP transit as a service ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs3 IP transit Lower tier ISPs customers of higher tier ISPs Critical for global connectivity Fee depends on traffic

4. Rade Stanojevic IP Transit costs ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs4 Affects almost every online entity IP transit cost-reduction technologies ISP peering CDN Multicast P2P localization Traffic shaping (P2P throttling) The less traffic on transit links the lower the cost

5. Rade Stanojevic Simple idea: Exploit the IP transit market properties: Price elasticity 95 th -percentile pricing CIPT: Cooperative IP transit ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs5 purchase in group and share the costs

6. Rade Stanojevic IP transit billing: two key properties ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs6 Price elasticity Peak based traffic metering: 95 th -percentile pricing Submodularity of the peak: peak(f) + peak(g) ≥ peak(f+g) Voxel pricing Source= https://www.voxel.net/ip-serviceshttps://www.voxel.net/ip-services (as accessed on June 2011)

7. Rade Stanojevic CIPT as a cooperative game ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs7 How to share the benefits of cooperation? Shapley value ‘Fair’, exists and unique for any cooperative game Individually rational (everyone better off in the coalition) Hard to calculate exactly for large number of players Monte-Carlo method for accurate estimation

8. Rade Stanojevic Part 2: Quantifying CIPT benefits ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs8

9. Rade Stanojevic From peering to transit traffic ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs9 ISP Transit traffic data: not available ISP Peering traffic data: available at some IXPs Peering traffic as a proxy for estimating transit traffic T transit (t) = cT peering (t) Validation using 4 medium-sized (academic) ISPs

10. Rade Stanojevic Obtaining peering traffic data ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs10 2. Transform images into numeric data on peering traffic 1. Peering traffic data of 264 ISPs from 6 IXPs Optical function/ character recognition Collection of mrtg images

11. Rade Stanojevic Expected cost savings in the range of 8%-56% ISP traffic volume distribution, peak-hour profile, nmb of transit providers, etc. Per ISP savings from $1K to few $100K per year Smallest ISPs save >90%, Largest ISPs save <10% Breakdown Around 1/3 of the savings come from the 95 th -pct The rest comes from the elastic pricing Check the paper for details… Country-wide coalitions ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs11

12. Rade Stanojevic Small (random) coalitions provide most of total attainable gains CIPT coalition size: low of diminishing returns ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs12

13. Rade Stanojevic Beyond gains sharing ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs13 Organizational embodiment Physical infrastructure Performance/SLAs Traffic confidentiality Transit providers and strategic issues

14. Rade Stanojevic Open problems CIPT coalition formation process? Shapley value too implicit. Simpler metrics? Tuángòu in other (price-elastic) domains? IP transport Wired/wireless access Mobile voice/txt/data ACM CoNEXT 201114CIPT: Using Tuángòu to Reduce IP Transit Costs

15. Rade Stanojevic Team-buying has potential for cutting IP transit bill Shapley value as a mechanism for cost sharing Unique dataset of traffic data from 264 ISPs with heterogeneous sizes and temporal properties Conclusions ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs15

16. Rade Stanojevic Backup slides ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs16

17. Rade Stanojevic Cooperative IP Transit (CIPT) ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs17 Tuángòu = team-buying coalitional arrangement for bulk buying of IP transit CIPT gains Per-Mbps price reduction thanks to elastic pricing and peak-based (95 th -pct) metering

18. Rade Stanojevic Price decay vs. traffic growth ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs18 Source= https://www.telegeographyhttps://www.telegeography Per-Mbps transit price decline vs. interdomain traffic growth

19. Rade Stanojevic Shapley value definition ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs19 Shapley value(i) ISP i's expected marginal contribution if the players join the coalition one at a time, in a uniformly random order N = number of players c(S)= cost of coalition S S(π, i) = set of players arrived in the system not later than i π = permutations of the set of players N i’s marginal contribution

20. Rade Stanojevic Monte Carlo method* We estimate the Shapley value as the average cost contribution over set π k of K randomly sampled arrival orders. Estimation accuracy K is the knob controlling the accuracy We use K = 1000 to keep the error under 1% (*) D. Liben-Nowell, A. Sharp, T. Wexler, K. Woods, “Computing Shapley Value in Cooperative Supermodular Games”, Preprint, 2010. R. Stanojevic, N. Laoutaris, P. Rodriguez, “On Economic Heavy Hitters: Shapley Value Analysis of the 95th-Percentile Pricing”, Proc. of ACM IMC 2010. Shapley value estimation ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs20

21. Rade Stanojevic Peering = 35-40% of the total traffic. Peering and transit follow very similar patterns. Peering-transit traffic similarity ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs21

22. Rade Stanojevic Absolute aggregate CIPT gains grow with IXP size (in terms of billed traffic) Relative aggregate CIPT gains decrease with IXP size Aggregate CIPT gains ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs22

23. Rade Stanojevic Per-partner CIPT gains ACM CoNEXT 2011CIPT: Using Tuángòu to Reduce IP Transit Costs23 Absolute individual CIPT gains grow with ISP size Relative individual CIPT gains decrease with ISP size There are large gains for all CIPT members